The potential of genetically engineered rodent models has accelerated demand for training procedures of behavioral tasks. Such training is generally time consuming and often shows large variability in learning speed between animals. To overcome these problems, we developed an efficient and stable training system for the two-alternative forced-choice (2AFC) visual stimulus detection task for freely behaving rodents. To facilitate the task learning, we introduced a spout-lever as the operandum and a three-step training program with four ingenuities: (1) a salient stimulus to draw passive attention, (2) a reward-guaranteed trial to keep motivation, (3) a behavior-corrective trial, and (4) switching from a reward-guaranteed trial to a nonguaranteed one to correct behavioral patterns. Our new training system realizes 1-week completion of the whole learning process, during which all rats were able to learn effortlessly the association between (1) lever-manipulation and reward and (2) visual stimulus and reward in a step-by-step manner. Thus, our new system provides an effective and stable training method for the 2AFC visual stimulus detection task. This method should help accelerate the move toward research bridging the visual functions measured in behavioral tasks and the contributing specific neurons/networks that are genetically manipulated or optically controlled.